Copyright in this work rests with the author. Please ensure that any reproduction or re-use is done in accordance with the relevant national copyright legislation.
ii
Approval
Name: Patrick Johnson
Degree: Master of Arts
Title: Cell towers and the ambient population: A spatial analysis of disaggregated property crime in Vancouver, BC
Examining Committee: Chair: Stephanie Wiley Assistant Professor
Martin Andresen Senior Supervisor Professor
Bryan Kinney Supervisor Associate Professor
Rémi Boivin External Examiner Professor School of Criminology University of Montreal
Date Defended/Approved: December 10, 2018
iii
Abstract
The current study employs a new measure of the ambient population, constructed using
cell tower location data from OpenCellID, to compare residential and ambient
population-based crime rates in Vancouver, BC. Five disaggregated property crime
types are examined at the dissemination area level. Findings demonstrate striking
differences in the spatial patterns of crime rates constructed using these two different
measures of the population at risk. Multivariate results from spatial error models also
highlight the substantial impact that the use of a theoretically-informed crime rate
denominator can have on Pseudo R2 values, variable retention, and trends in significant
relationships. Implications for theory testing and policy are discussed.
Keywords: ambient population; OpenCellID; population at risk; property crime; spatial
analysis; Vancouver
iv
Acknowledgements
I owe a number of people my gratitude for the roles they played in helping me get
to this point. First, thank you to my senior supervisor, Dr. Martin Andresen, for your
guidance over the course of this project and for the opportunity to work in the ICURS lab.
Thank you as well to my other committee members, Dr. Bryan Kinney and Dr. Rémi
Boivin, for taking the time to read my thesis and provide criticism. I also appreciate the
professors, classmates, and all the other people that have made my (many) years at
SFU interesting. I’ve learned a lot here, much of it outside the realm of academia.
Thank you as well to my friends, who listened to my various complaints
throughout the thesis process; you helped more than you might realize. Lastly, thank you
to my family. Without your support I’m not sure I would have finished my thesis. The
advice and encouragement – and meals! – you provided helped me immensely. I’m very
lucky to have a father, mother, and sister like you.
v
Table of Contents
Approval .......................................................................................................................... ii
Abstract .......................................................................................................................... iii
Acknowledgements ........................................................................................................ iv
Table of Contents ............................................................................................................ v
List of Tables ................................................................................................................. vii
List of Figures................................................................................................................ viii
List of Acronyms ............................................................................................................. ix
Routine activity theory differs from social disorganization theory in that the focus
is the criminal event, not neighbourhood-level processes (Andresen, 2014). Cohen and
Felson (1979) developed routine activity theory to explain a sociological paradox: even
though socioeconomic conditions in the post-war United States had improved, crime
rates were rising substantially. Despite important decreases in unemployment and the
number of people living below the legally-defined poverty level, as well as increases in
education and median family income, crime rates for various property and violent crime
types had risen between 164 and 263 percent between 1960 and 1975 (Cohen &
Felson, 1979).
Cohen and Felson (1979) theorized that changes in people’s routine activities
could explain this paradox. Routine activities are defined as “any recurrent and prevalent
activities which provide for basic population and individual needs, whatever their
biological or cultural origin” (Cohen & Felson, 1979, p. 593). After the Second World War
people had more disposable income, college enrollment increased, and there was
greater female labour force participation (Cohen & Felson, 1979). These societal
changes altered people’s routine activities. More people were out of the home which put
them at greater risk of victimization.
Routine activity theory is concerned with criminal opportunity. A criminal event is
explained as the result of the spatiotemporal convergence of three factors: a motivated
offender, a suitable target, and a lack of guardianship (Cohen & Felson, 1979). Worth
noting, is that routine activity theory only explains direct-contact predatory violations,
such as homicide, assault, break-and-enter, and theft (Andresen, 2014). Cohen and
Felson’s (1979) analysis generally supported their theory at the national level. Like social
disorganization theory, subsequent research has demonstrated the utility of routine
activity theory as a theoretical framework for analyzing crime spatially (see Andresen,
2006a; Murray & Swatt, 2013; Nogueira de Melo et al., 2017). Specifically, this
theoretical perspective proved useful in terms of variable selection, accounting for
6
temporal patterns of population movements and crime, and offering insights into variable
significance trends.
The current study incorporates both social disorganization theory and routine
activity theory to capture neighbourhood and event-level factors associated with crime.
Prior studies have used both frameworks to conduct analyses at a variety of spatial
scales, from street face blocks (Rice and Smith, 2002; Smith, Frazee, & Davison, 2000)
to census tracts (Andresen, 2006a; Willits, Broidy, & Denman, 2013). By and large, prior
research has supported the integration of both theories. For this reason, both theories
were used to inform the selection of independent variables to increase the predictive
power of the models (Andresen, 2006a). Still, it should be noted that in some cases
these theories may produce conflicting expectations regarding the relationship between
an independent variable and crime (Andresen, 2006a). For example, social
disorganization theory would predict that median dwelling values would be negatively
associated with crime rates, since areas with more expensive homes have higher
socioeconomic status. By contrast, under routine activity theory the relationship would
be positive, since these areas would likely have more suitable targets. The researcher
must simply be aware of this possibility when using both theoretical frameworks.
2.3. A Brief History of the Use of Alternative Denominators
Alternative denominators have been used in the study of crime at least as far
back as 1938, when Lottier examined state-by-state differences for a variety of crime
types. When calculating auto theft rates, Lottier (1938) used the number of automobiles
registered in the state as the population at risk. While Lottier (1938) did not discuss the
reasoning behind his decision to use this denominator, doing so would suggest that he
felt this measure would better capture environmental risk for this crime type than the
residential population. Indeed, with lower rates of car ownership during the 1930s, a rate
based on the residential population would likely have been low and would not have
provided an accurate indication of the risk facing automobile owners.
The first study that examined and compared crime rates with alternative
denominators to traditional ones was conducted by Boggs (1965). Boggs (1965)
calculated correlations between residential population-based crime rates and crime-
specific rates based on environmental opportunities for St. Louis census tracts. Boggs
7
(1965) pointed out that crime rates based on the residential population may lead to
spuriously high rates for central business districts, which often have few residents, but
“large numbers of such targets as merchandise on display, untended parked cars on
lots, people on the streets, money in circulation, and the like” (p. 900). When the large
number of criminal opportunities offered by central business districts is considered, it
may very well be that environmental risk is lower in these areas, relative to other parts of
the city.
To provide better indications of environmental opportunities and risk, Boggs
(1965) used a variety of clever alternative denominators, including the following:
Table 2.1. Boggs’ (1965) alternative denominators
Crime Type Alternative Denominator
Auto theft Space devoted to parking
Highway (street) robbery Square footage of streets
Homicide and aggravated assault Pairs of persons
Non-residential burglary Business-residential land use ratio
Adapted from Boggs (1965)
While some of Boggs’ (1965) crime-specific denominators provided questionable
representations of environmental opportunity (e.g. square footage of streets), they were
an important first step in the use of alternative crime rate denominators.
Boggs (1965) found that some of the traditional, residential population-based
crime rates were highly correlated with their alternative counterparts. For example,
criminal homicide and aggravated assault, forcible rape, and residential day burglary all
had rank order correlations of 0.997, 0.969., and 0.924, respectively. This finding
suggests that for these crime types, the alternative denominators were likely not of any
particular value; the residential population was most likely capturing the population at
risk. Other crime types had lower correlation coefficients between the traditional and
crime-specific rates, such as auto theft for joy riding and business robbery. Interestingly,
three crime types had negative correlations between the standard and alternative rates:
non-residential night burglary, non-residential day burglary, and grand larceny. Boggs
(1965) investigated non-residential night burglary further, comparing the rankings of both
8
types of rates across St. Louis census tracts. She found that the census tracts with the
highest rankings for traditional rates ranked near the bottom for crime-specific ones.
Lending support to her claim about central business districts having spuriously high
crime rates, Boggs (1965) found that these census tracts often had low residential
populations and high ratios of business to residential land use. When the number of
criminal opportunities in these census tracts was accounted for with an alternative
denominator, it became apparent that residential population-based rates were vastly
overstating risk in these areas. Overall, Boggs’ (1965) seminal work suggested that for
some crime types, the residential population may be an inappropriate and misleading
denominator.
Subsequent research conducted by Skogan (1976) reinforced Boggs’ (1965)
findings of potentially important differences between residential population-based crime
rates and alternative ones. In one of his analyses, Skogan (1976) compared rates of
motor vehicle theft in several large American cities per 1,000 residents to rates per 1,000
vehicles. He found that while New York City ranked quite low using the traditional rate
(12 motor vehicle thefts per 1,000 residents), it ranked first amongst the cities studied
when the alternative rate was used (53 motor vehicle thefts per 1,000 vehicles). This
finding underscores the importance of selecting “meaningful denominators, to analyze
victimization experiences in light of the exposure of potential victims to risk” (Skogan,
1976, p. 172). Skogan (1976) noted that fewer people own cars in New York City,
meaning that the motor vehicle theft rate with number of vehicles as the denominator
likely provides a better indication of risk for vehicle owners.
In contrast to the findings of Boggs (1965) and Skogan (1976), later work by
Cohen, Kaufman, and Gottfredson (1985) suggested that concerns about the accuracy
of residential population-based crime rates may be unwarranted. Cohen et al. (1985)
found that when traditional and alternative rates for burglary and auto theft were
compared, they were quite similar. Moreover, the traditional rates consistently provided
better forecasts than the alternative rates. It is worth noting that Cohen et al. (1985) only
examined two crime types; it was premature of them to argue that it does not matter
whether traditional or alternative denominators are used in the calculation of crime rates.
Returning to Boggs (1965), even though some alternative rates did not provide new or
different information, others certainly did.
9
In the case of residential burglary, it is not all that surprising that Cohen et al.
(1985) found that crime rates calculated using traditional and alternative denominators
were highly correlated. As pointed out by Andresen (2014), where there are higher
residential populations, there are usually a greater number of households (the alternative
denominator used by Cohen et al. (1985) for residential burglary). Had Cohen et al.
(1985) examined a different crime type, where the residential population is not likely to
be as highly correlated with an alternative population at risk denominator (e.g.
commercial burglary), they may have found more important distinctions between crime
rates calculated using traditional and alternative denominators.
The relative paucity of early studies making use of alternative denominators in
the calculation of crime rates has been attributed to the high cost and difficulty of
obtaining these measures (Harries, 1991). Clarke (1984) also pointed out the conceptual
difficulties associated with defining the population at risk for certain crime types.
Returning to Boggs (1965), the relevance of the square footage of streets to the risk of,
or opportunity for highway robbery is debatable. Nevertheless, this measure recognizes
that the residential population does not accurately capture the population at risk for this
crime type. Indeed, the crime committed in a given spatial unit “is not limited to crimes
committed by residents” (Gibbs & Erickson, 1976, p. 606). As part of their routine
activities, people move around, and are still at risk of being victimized outside of the area
in which they live. Until recently, it would have been difficult to generate an estimate of
the number of people in a particular area engaged in their day-to-day activities; Boggs’
(1965) use of the square footage of streets represents an early attempt. Fortunately,
recent technological advances have permitted the development of new measures
designed to capture this population at risk, commonly referred to as the ambient
population.
10
Chapter 3. The Ambient Population
There is a small but growing literature in the social sciences1 on the use of the
ambient population as an alternative denominator. Researchers have calculated this
measure in a variety of innovative ways, including 24-hour average population estimates
from LandScan Global Population Database (Andresen, 2006b, 2011; Piza & Gilchrist,
2018), Twitter messages (Hipp et al., 2018; Kounadi et al., 2018; Malleson & Andresen,
2015a), pedestrian movement models (Chainey & Desyllas, 2008), and transportation
survey data (Boivin, 2018; Felson & Boivin, 2015). These and other measures have
been used to examine many different crime types, such as snatch-and-run offenses
(Hanaoka, 2018), stranger assaults (Boivin, 2013), and automotive theft (Andresen,
2006b). Overall, the literature consistently demonstrates the value of considering this
alternative denominator in crime analysis.
3.1. Non-Inferential Findings
In terms of the measures themselves, the literature consistently demonstrates
important differences between residential and ambient populations. Using a 24-hour
average population estimate from LandScan Global Population Database2 as a measure
of the ambient population in their analysis of violent crime in Vancouver, Andresen and
Jenion (2010) found that at the enumeration area level the ambient population had a
much wider range than the residential population. Specifically, the residential population
for enumeration areas ranged between zero and 1,832, while the ambient population
ranged between zero and 8,257 (Andresen & Jenion, 2010). Other studies have also
found wider ranges in various measures of the ambient population, compared to the
1 The ambient population has seen use in other fields, notably the computer sciences. While some of these applications have considered crime (see Bogomolov et al., 2014; Gerber, 2014; Traunmueller, Quattrone, & Capra, 2014), they were not largely informed by criminological theory and will not be covered in this literature review.
2 This measure of the ambient population estimates the number of people in a square kilometer spatial unit at any given time of day or year. It is calculated using census population data, road proximity, land surface slope, land cover, and nighttime lights. See Andresen (2006b) and Dobson et al. (2003) for more detailed discussions of this measure.
11
residential population (see Boivin, 2013; Malleson & Andresen, 2016; Mburu & Helbich,
2016). Andresen and Jenion (2010) noted that the ambient population for Vancouver
was higher than the residential population (547,000 compared to 514,000), indicative of
commuters coming in from the suburbs. When only the residential population is used as
a population denominator for a crime rate, commuters are not considered as part of the
population at risk.
Research also demonstrates greater clustering of ambient populations,
compared to residential populations. In their analysis of a variety of crime types across
Greater London census administrative areas, Mburu and Helbich (2016) used a temporal
weighting scheme that incorporated both residential population and workday population
measures to provide an estimate of the ambient population. They found that the ambient
population was more clustered in the city centre, compared to the residential population
(Mburu & Helbich, 2016). Malleson and Andresen (2016), as well as Andresen and
Jenion (2010), reported similar findings in London and Vancouver, respectively. Notably,
all four of the ambient population measures evaluated by Malleson and Andresen (2016)
(census workday population, geo-located Twitter messages, mobile telephone activity
counts, and Population 24/7 population estimates) demonstrated clustering in London’s
city centre. These findings speak to Boggs’ (1965) assertion that low residential
populations in central business districts could lead to spuriously high crime rates in these
areas and inaccurate portrayals of risk.
Before moving on to the use of the ambient population in crime rate calculations
there is one final point worth noting: studies making use of ambient population measures
have routinely found low correlations between residential and ambient populations (see
Helbich, 2016). These findings speak to the important differences between the
residential and ambient population. Given these differences, it is clear that these two
measures of the population at risk are not interchangeable (Andresen & Jenion, 2010).
3.1.1. Ambient Population-Based Crime Rates
While it has generally been found that correlations between residential and
ambient population measures are low, two studies did find that crime rates making use
of both of these population denominators were highly correlated. Andresen (2006b)
12
analyzed automotive theft, break-and-enter, and aggregate violent crime in Vancouver.
He found that for all three crime types the residential and ambient population-based
crime rates were all highly correlated. In a later study, also set in Vancouver but using
more recent crime data, Andresen (2011) again found that aggregate violent crime rates
using residential and ambient population denominators were highly correlated at both
the census tract and dissemination area levels. Interestingly, when earlier census
boundaries were used the results for dissemination/enumeration areas changed
dramatically. Specifically, the r values went from 0.801 to 0.095. This finding suggests
that the high correlations between residential and ambient population-based crime rates
may not be universal and that further research on the subject is required.
At the municipal level, two studies have examined differences between crime
rates calculated using residential and ambient population measures. Andresen (2010)
used both LandScan Global Population Database data and census survey data on
commuting trips to estimate ambient populations in Greater Vancouver Regional District
municipalities. He found that while municipalities with the highest resident-based crime
rates also had the highest ambient-based rates, there was more variation between the
lower ranked municipalities. Subsequent research conducted by Stults and Hasbrouck
(2015) employed US census data on commuting to construct alternative crime rate
denominators based on daytime population changes in large American cities. They
found considerable changes in cities’ crime rate rankings, depending on which
denominator was used. Clearly, the population at risk denominator matters when
considering municipal crime rates.
Other studies employing ambient population measures have examined the
spatial patterning of crime rates, both visually and statistically. One recurrent finding is
that when ambient population measures are used as denominators, crime rates in city
centres decrease. This finding speaks to Boggs’ (1965) early work and confirms the
importance of selecting appropriate denominators to account for the population at risk.
Using LandScan Global Population Database data as an ambient population
denominator, Andresen (2011) found that aggregate violent crime rates dropped in
Vancouver’s downtown area. This finding is intuitive; city centres and downtown areas
bring together large numbers of people engaging in their routine activities. When this
larger ambient population was accounted for, it makes sense that rates of violent crime
in these areas would have decreased. By definition, violent crime requires people to
13
come together in time and space. The results of Andresen and Brantingham (2007),
along with those of Mburu and Helbich (2016) echo these findings. When commuting
populations in London were accounted for, violent crime rates in central authority
districts, such as the City of London and Westminster, fell drastically (Mburu & Helbich,
2016). Though not a crime rate per se, in their report on Vancouver hotspots of crime
Andresen and Brantingham (2007) controlled for the ambient population using dual
kernel density maps. They found that both violent and property crime hotspots in
Vancouver’s downtown decreased in their intensity.
Other spatial analyses making use of ambient population measures have, more
generally, identified shifts in crime rate clusters as well as new clusters, depending on
the population denominator used. Mapping local indicators of spatial association (LISA)
for aggregate violent crime rates, Andresen (2011) found consistent shifting of crime hot
spots toward Vancouver’s downtown peninsula, when ambient rates were compared to
residential ones. This finding was consistent at both the census tract and dissemination
area levels of analysis. Later work by Malleson and Andresen (2016) identified new
hotspots when significant Getis-Ord GI* clusters of ambient population-based rates of
theft from persons offenses were mapped. These clusters, where the risk of being a
victim of a theft from persons offense was higher, would not have been identified using
only a residential population-based crime rate. Another study by Malleson and Andresen
(2015b) also employed the Getis-Ord GI* statistic, that measures spatial clustering. They
found that there was statistically significant clustering of residential population-based
violent crime rates in Leeds’ city centre. However, when the ambient population was
accounted for, using geo-located Twitter messages, these clusters became insignificant.
Despite a high volume of violent crime events in this area, the risk of being a victim of
violent crime was not significantly higher when the ambient population was taken into
account.
While the above studies have focused exclusively on the spatial patterning of
ambient population-based crime rates, Malleson and Andresen (2015a) employed a
novel approach to examine spatiotemporal hotspots of robbery and theft from persons in
Leeds. One of their findings was the identification of a cluster near the University of
Leeds campus, from 21:00 on Saturdays to 2:00 on Sundays. The authors did
acknowledge several limitations of their data, such as the fact that the social media data
used to estimate the ambient population did not cover the same time period as the crime
14
data. Still, this study represents an innovative step in studying and understanding
spatiotemporal trends in ambient population-based crime rates. Malleson and
Andresen’s (2015a) study, along with the ones discussed above, highlight the important
differences between crime rates calculated using the residential and ambient
populations. They also demonstrate the valuable information that can be gained when
this alternative denominator is used. This body of research seriously brings into question
Cohen et al.’s (1985) assertion that it does not matter which population denominator is
used.
3.1.2. Other Non-Inferential Work
Before moving on to inferential research, I will briefly discuss two other studies
that employed ambient population measures exclusively. Felson and Boivin (2015) used
transportation survey data to capture the number of daily visitors in census tracts in a
large Eastern Canadian city. They found that various visitor types were strongly linked to
aggregate property and violent crime (Felson & Boivin, 2015). Kurland, Johnson, and
Tilley (2014) compared rates of violent crime and theft and handling offenses around a
UK stadium, using LandScan Global Population Database ambient population data and
match/event ticket sales to provide population at risk estimates for regular days and
match/event days, respectively. By using these two ambient population measures, the
authors did not rely on the residential population at all. One interesting finding was that
although counts of theft and handling offenses were much higher on match and event
days, compared to days when neither occurred, the rates of these offenses where
significantly lower on match and event days (Kurland et al., 2014). When the increased
population in the area surrounding the stadium on match and event days was accounted
for, it was determined that the risk of being a victim of these types of offenses was
actually lower than on days with no matches or events. In their use of these ambient
population measures, both Felson and Boivin (2015) and Kurland et al. (2014)
recognized the importance of considering the population at risk.
3.2. Inferential Findings
Overall, the ambient population has seen more use in a descriptive context. Still,
there are several studies that have used this measure inferentially. When assessing the
15
relationship between the two measures, Andresen and Jenion (2010) found that the
residential population was a poor predictor of the ambient population, with an adjusted-
R2 value of .287. The residential population-based violent crime rate performed even
more poorly as a predictor of the ambient-based rate (adjusted-R2 = 0.007). Given these
findings, Andresen and Jenion (2010) concluded that the two population measures are
likely not substitutable.
3.2.1. Use of the Ambient Population as an Independent Variable
While the primary focus of his research was on the location quotient, a measure
of a region’s specialization in a particular crime type, Andresen (2007) incorporated the
ambient population into his analysis. He found that the while the ambient population was
positively associated with a census tract’s specialization in automotive theft, it was
negatively associated with break-and-enter specialization. Andresen (2007)
hypothesized that in the case of automotive theft, larger ambient populations mean more
vehicles and more potential targets, leading to specialization in this crime type. As for
break-and-enter, Andresen (2007) suggested that larger ambient populations provide
guardianship, decreasing specialization. The ambient population was not found to be
associated with specialization in violent crime. So, while the ambient population may be
a better measure of the population at risk for violent crime, it is not associated with a
census tract’s specialization in violent crimes.
Within the last five years there has been a growing number of multivariate
analyses conducted using ambient population measures as independent variables. In
the context of commuting, Boivin (2013) used the number of workers in Montreal census
tracts to estimate the ambient population. Although the residential population was found
to have a significant positive effect on both domestic violence and burglaries, no such
relationship existed for stranger assaults. The ambient population, however, emerged as
a strong, significant predictor of the number of stranger assaults. Stults and Hasbrouck
(2015) used commuting data as well to examine crime rate estimates at the municipal
level. Daily commuting rates were found to be a strong predictor of overall crime rates in
American cities.
As discussed above, the number of daily visitors to a census tract has also been
used to capture the ambient population. Both Boivin (2018) and Boivin and Felson
16
(2018) calculated census tract visitors using transportation survey data. Boivin and
Felson (2018) found that an increase in visitors was associated with both more visitors
and residents being charged with a crime in that census tract. By contrast, Boivin’s
(2018) research suggested that the relationship between crime and visiting populations
is more ambiguous. Using geographically weighted regression, Boivin (2018) found that
larger visiting populations were associated with higher levels of crime. However, for
some visit types, the relationship with crime was negative. This finding suggests that in
some cases, larger populations may provide guardianship (Boivin, 2018).
The issue of guardianship is further muddied by Hanaoka’s (2018) research in
Osaka on snatch-and-run offenses. Hanaoka (2018) used average hourly weekday
ambient population counts based on cell phone users who had their ‘Auto GPS’’ function
enabled. Hanaoka (2018) found that while elevated ambient population levels were
associated with more snatch-and-run offenses at night, the opposite was true during
daytime. This research highlights the importance of considering temporal trends when
conducting spatial analysis.
Two final studies worth mentioning in this section were conducted by Hipp et al.
(2018) and Kadar and Pletikosa (2018). In their assessment of routine activity theory and
crime pattern theory, Hipp et al. (2018) found that their temporal ambient population
measure, geolocated Twitter data, was useful in explaining crime at the city block level.
At the census tract level, Kadar and Pletikosa’s (2018) human mobility measure was
found to increase absolute R2 metrics for their models. This measure was calculated
using data on subway and taxi usage in New York City, as well as data from Foursquare.
Taken together, the above studies demonstrate the value of including ambient
population measures in multivariate analyses of crime. Including this measure as an
independent variable in regression models can provide insights that are missed when
only the residential population is considered.
3.2.2. Use of Ambient Population-Based Crime Rates as Dependent Variables
Only three studies have employed ambient population-based crime rates as
dependent variables. A recurrent theme amongst these three studies concerns model
goodness of fit. Specifically, regression models that employ ambient population-based
17
crime rates as dependent variables consistently demonstrates better goodness of fit than
their residential-based counterparts. Andresen (2006b) found that for both automotive
theft and break-and-enter the Pseudo R2 was substantially higher for the models using
ambient population-based crime rates. For example, the Pseudo R2 for automotive theft
jumped from 0.538 to 0.723. Additionally, the ambient rate regression model for
automotive theft retained twice the number of variables compared to the residential rate
model.
Somewhat surprisingly, Andresen (2006b) found that for aggregate violent crime,
the model with the residential population-based rate provided superior goodness of fit.
This result was unexpected, given that the ambient population would seem to measure
the population at risk for these types of crimes best. Nevertheless, later research on
aggregate violent crime rates by Andresen and Brantingham (2007) and Andresen
(2011) found that models with ambient-based rates had higher Pseudo R2 values than
those using residential-based rates. Andresen’s (2011) results were consistent at both
the census tract and dissemination area levels, making them all the more credible.
Andresen and Brantingham (2007) found that the ambient population-based property
crime rate provided better goodness of fit as well.
It is beyond the scope of this paper to list and discuss all the independent
variables associated with ambient population-based crime rates in the studies discussed
above. Still, it is worth mentioning that variables linked to both routine activity theory and
social disorganization theory have consistently been associated with ambient population-
based crime rates (see Andresen, 2006b, 2011; Andresen & Brantingham, 2007). For
example, ambient population-based rates of violent crime have been positively
associated with population change (Andresen, 2006b) and recent movers (Andresen &
Brantingham, 2007). These two indicators relate to Shaw and McKay’s (1942) construct
of the physical status of a neighbourhood. In terms of the routine activity theory concept
of a suitable target (Cohen & Felson, 1979), Andresen & Brantingham (2007) found that
both percentages of those receiving government assistance and average dwelling
values were negatively associated with ambient population-based rates of aggregate
property crime. Individuals living in these areas likely have fewer possessions that would
be considered suitable targets by offenders.
18
It should be noted that these three studies were all conducted in Vancouver;
replication in other settings is necessary before any sweeping theoretical assertions are
made. In terms of similarities between residential and ambient population-based violent
crime rate models, Andresen (2011) found that both the signs and magnitudes of the
independent variables were comparable. Andresen (2011) also noted more variable
retention for the ambient rate model at the census tract level. In light of these findings,
Andresen (2011) concluded that the ambient population is “likely better than the
residential population when analyzing violent crime” (p. 209).
3.3. Summary
Taken together, the above literature on alternative denominators and the ambient
population supports Boivin’s (2018) assertion that “other populations matter” (p. 83).
Whether used to map crime rate hotspots or as an independent variable in a spatial
regression model, the ambient population consistently provides important information
that would have been missed had only the residential population been considered. This
is not to say that the residential population is useless. Rather, it should not simply be
assumed that it best captures the population at risk for every crime type.
Still, the difficulties in operationalizing the ambient population should be
acknowledged. As Malleson and Andresen (2015a) pointed out, Twitter data likely
contains omissions and will almost certainly over and under-represent different groups.
The number of workers in a census tract (Boivin, 2013) also provides a questionable
representation of the ambient population; this measure misses groups like youth,
students, and the unemployed. Finally, mobile data like the kind used in Hanaoka’s
(2018) study rely on users having a GPS function on their phone enabled. Because of
the biases inherent to most, if not all, ambient population measures, they should be used
alongside the residential population.
To conclude this brief summary, a major gap in the literature on the ambient
population should be mentioned. Inferential research that uses ambient population-
based crime rates as dependent variables is extremely limited (Andresen, 2006b, 2011;
Andresen & Brantingham, 2007). The studies that have done so have typically found that
regression models for ambient population-based rates have better goodness-of-fit than
residential ones. Of these three studies, only one was conducted at the dissemination
19
area level (Andresen, 2011); the others examined census tracts. Larger spatial units like
census tracts may mask heterogeneity within (Andresen & Malleson, 2013). These three
prior studies also all employed the same ambient population measure obtained from
LandScan Global Population Database (Andresen 2006b, 2011; Andresen &
Brantingham, 2007). Lastly, the work of Andresen (2006b, 2011) and Andresen and
Brantingham (2007b) used aggregate measures of crime. Research by Andresen and
Linning (2012) found that the use of aggregated crime types in spatial analysis often
masks important spatial patterns. They conclude that the use of aggregated crime types
in spatial analysis is inappropriate (Andresen & Linning, 2012). These limitations of the
prior inferential research using ambient population-based crime rates justifies the current
study and its design.
20
Chapter 4. Data and Methods
The current study examines and compares residential and ambient population-
based crime rates for disaggregated property crime types in Vancouver, British
Columbia at the dissemination area level. While descriptive findings and maps will be
presented, the study is primarily inferential. Spatial error models for both residential and
ambient population-based rates are produced and compared. Model variables are
informed by both social disorganization theory and routine activity theory. The ambient
population measure was constructed using open source cell tower location data from
OpenCellID (https://opencellid.org/).
4.1. Data
With over 631,000 residents, the City of Vancouver is the most populous
municipality in the province of British Columbia, and the entire metropolitan area is the
third-largest in Canada. Over the years, Vancouver has experienced consistent
population growth, with a moderate 4.6 percent increase from 2011 to 2016. Going
further back to 1991, Vancouver’s residential population has grown a full 33.8 percent. In
terms of landmass, the City of Vancouver occupies 114.97 square kilometers on the
western part of the Burrard Peninsula. As a seaport on the Pacific Ocean, Vancouver is
crucial to Canada’s international trade. By tonnage, the Vancouver Fraser Port Authority
is the largest port in Canada, and the third-largest in the Americas. Average income for
individuals in 2016 was CAN$ 50,317, slightly higher than the national average of CAN$
47,487. Vancouver is also noted for being part of one of the most ethnically diverse
metropolitan areas in Canada.3
As noted in the 2016 census data, the City of Vancouver is the eighth-largest
municipality in Canada. Interestingly, the total number of Criminal Code offenses per
100,000 residents in 2016 (8,243) was much higher than Canada’s two largest
3 All statistics in this paragraph obtained from “Vancouver” (n.d.).
21
municipalities, Toronto (3,741) and Montreal (4,351). Property crime, the focus of the
current study, paints a similar picture. Total property crime per 100,000 Vancouver
residents in 2016 was 6,172, compared to 2,314 for Toronto and 2,675 for Montreal. It is
worth noting that Vancouver’s total property crime rate is nearly double that of Canada
as a whole (3,225), suggesting the need for further study of this phenomenon.4
The dataset used in this study consists of three data sources: property crime
data from Vancouver’s municipal police force (VPD), Statistics Canada census data, and
open source cell tower location data from OpenCellID. The year 2016 was chosen for
the VPD data to correspond with Canada’s most recent census. As noted by Andresen
(2006a), failure to use crime data from the same year as sociodemographic and
socioeconomic indicators may limit interpretation of the findings. It may be that
relationships between crime and census variables of different years are spurious. These
three data sources are discussed below.
4.1.1. Crime Data
The 2016 property crime data from the VPD are made up of seven crime types:
commercial break-and-enter, residential break-and-enter, mischief, theft from vehicle,
theft of vehicle, theft of bicycle, and other theft. These data are in their raw, count form
and together make up total property crime in Vancouver. In 2016, there were 2436
commercial break-and-enters, 2994 residential break-and-enters, 3938 counts of
mischief, 8870 thefts from vehicle, 1288 thefts of vehicle, 2405 thefts of bicycle, and
5708 other thefts. Taken together, there were 27, 639 property crimes reported to the
VPD in 2016.
These data, from 2003 to present, are publicly available through the City of
Vancouver’s open data catalogue (http://vancouver.ca/your-government/open-data-
catalogue.aspx). Criminal events are geocoded at the hundred block level. The data
represent police calls for service (CFS). Andresen (2006a) points out that CFS are, in
fact, a proxy for criminal activity, because a charge may or may not be laid as a result of
the call. Still, an advantage of this data source is that it provides a better indication of
police activity than Statistics Canada data.
4 All statistics in this paragraph obtained from Statistics Canada (2017).
22
There are documented issues with the use of official crime data, namely the dark
figure of crime. For instance, MacDonald (2001) found that individuals not in the labour
market were less likely to report a property crime. For the current study, this finding
suggests that there could be systematic variations in reporting levels across
dissemination areas, depending on their labour force participation rate. Indeed,
unemployment in Vancouver dissemination areas ranges from 0 percent to 30.3 percent.
Many of these dissemination areas with higher levels of unemployment are concentrated
in Vancouver’s Downtown Eastside neighbourhood. While little, if anything, can be done
to address this data limitation, it is worth acknowledging.
4.1.2. OpenCellID
The OpenCellID data source was used to create an ambient population measure
that could be used as an alternative crime rate denominator to the residential population.
OpenCellID describes itself as “the world’s largest collaborative community project that
collects GPS positions of cell towers” (OpenCellID, 2018). Users typically join to obtain
location services information on their mobile devices without relying on GPS, as well as
to research cell tower coverage. It should be noted that the data actually represent cells
in cellular networks. Individual cells are serviced by base transceiver stations, or BTS,
that use antennae fixed to cell towers to provide network coverage. Often, there are
multiple antennae from multiple providers on a single tower. The size of the cell service
area depends on a variety of factors, such as the number of users and the
characteristics of the surrounding environment (e.g. topography, weather).
The cell location data are open source and are directly downloadable from
opencellid.org in a .gz ZIP file. The file used in the current study was downloaded on
September 14th, 2017. The data are cumulative, with user-identified cells being added to
the database over time. Using a spatial join function in ArcMap 10.3 cell locations were
geocoded (100% hit rate) to dissemination areas in Vancouver and its surrounding
municipalities (Metro Vancouver). 19215 unique cells were identified in the City of
Vancouver itself, with dissemination area cell counts ranging from zero to 732. Not
surprisingly, the dissemination area with the highest cell count was located in
Vancouver’s downtown core. Cells also clustered at major population centers and along
transportation corridors. These finding leads into the justification for using cell counts to
create an ambient population measure.
23
The Canadian Wireless Telecommunications Association (CWTA) notes that as
of December 2017, approximately 90% of Canada’s population subscribed to mobile
services (CWTA, n.d.). Because nearly everyone has a cell phone, it is argued that cell
data from OpenCellID can be used to construct a new ambient population measure.
Areas with larger ambient populations (such as Vancouver’s downtown core) require
more cells to support these users. Therefore, areas with greater concentrations of cells
likely have higher ambient populations. To the researcher’s knowledge, no prior studies
have made use of this data source to study crime.5
To create this ambient population measure, the residential population of
Vancouver and its surrounding municipalities (Metro Vancouver) was proportionately
redistributed based on dissemination area cell counts. For the purposes of this study,
Metro Vancouver was considered a relatively closed system; the residential populations
of Vancouver’s surrounding municipalities were taken into consideration to account for
daily population flows of people engaged in their routine activities. In cases where
dissemination areas had zero cells, the average calculated ambient population from the
nearest spatial neighbours was used (Queen’s contiguity 1). At approximately 800,000
persons, the total calculated ambient population for Vancouver is 27% larger than its
residential population.
5 OpenCellID has been used as a data source in research on wireless networks and telecommunications (see Frank, Mannor, & Precup, 2013; Lee, Shih, & Chen, 2013; Xie, Heegaard, & Jiang, 2017). Only one study in the social sciences field using OpenCellID as a data source was identified (Hodler & Raschky, 2017). This study examined the relationship between ethnic groups and mobile phone infrastructure in Africa.
24
Figure 4.1. Vancouver’s residential population
Figure 4.2. Vancouver’s ambient population
25
Figure 4.3. Percent change, residential to ambient population for Vancouver
Figures 4.1., 4.2., and 4.3. highlight the striking differences in the ranges and
spatial patterning of Vancouver’s residential and ambient populations. While the
residential population of Vancouver’s dissemination areas ranges from 68 to 8778, the
ambient population ranges from 40.94 to 29970.7. In terms of percent change, the
difference between the residential population and the ambient population in Vancouver’s
dissemination areas ranged from approximately -95 percent all the way to 2266 percent.
These figures depict ambient population clustering in Vancouver’s downtown core. As
discussed in the literature review, using the residential population as a crime rate
denominator in this area could lead to spuriously high crime rates.
To conclude this subsection, three limitations of this ambient population measure
should be noted. First, the locations of the cells are averaged based on multiple
measurements from OpenCellID users, meaning that their recorded locations may differ
slightly from their actual locations. Second, and somewhat obviously, the data from
OpenCellID is user-generated. As noted by Malleson and Andresen (2015a) in their
study that estimated the ambient population using geo-located Twitter messages, there
26
may be omissions and biases. For instance, homeless and poorer populations may have
lower rates of mobile service subscription. These populations may be under-represented
in the current study’s ambient population measure. Lastly, because the data are
cumulative, seasonal and event-driven population changes cannot be accounted for.6
4.1.3. Census Data
The 2016 Statistics Canada census data used in this study were retrieved from
the Canadian Socioeconomic Information Management (CANSIM) database. The data
retrieved are at the dissemination area level, the smallest census unit in Canada with
sociodemographic and socioeconomic data. In 2016 there were 991 dissemination areas
in Vancouver. Of these 991 dissemination areas, 13 were excluded from this analysis
because, for confidentiality reasons (due to low residential population counts), Statistics
Canada did not release sociodemographic and socioeconomic data.
Seventeen sociodemographic and socioeconomic indicators were selected from
the 2016 Statistics Canada data for use as independent variables in this analysis. Most
of these variables were converted to percentages, for ease of interpretation. Variables
were chosen based on their relevance to Shaw and McKay’s (1942) social
disorganization theory and Cohen and Felson’s (1979) routine activity theory, the two
theoretical perspectives that underpin the current study. As discussed above, population
composition is a key construct in in social disorganization theory (Shaw & McKay, 1942).
Percentages of Aboriginals, visible minorities, immigrants, and ethnic heterogeneity were
chosen to reflect this construct. The ethnic heterogeneity variable was calculated from
census data on ethnic origins. Using this data, scores on the Blau (1977) index were
generated. A score of zero indicates no mix of ethnic groups (i.e. ethnic homogeneity),
whereas a score of one hundred indicates an even ethnic mix (i.e. perfect ethnic
heterogeneity).
Percentages of recent immigrants, people who moved into the dissemination
area within the last year, and rented households were used to capture population
Rented households also have relevance when it comes to the routine activity theory
6 For a study on the influence of weather and seasonality on pedestrian traffic volumes see Aultman-Hall, Lane, and Lambert (2009).
27
concept of guardianship; renters are expected to engage in more activities away from
home (Andresen, 2006a). To measure economic status, median income along with the
percentages of unemployment, government assistance, low income designation,
subsidized housing, housing under major repair, and post secondary education levels
were used. Family disruption (Sampson & Groves, 1989) was measured using the
percentage of lone parents. Lastly, the number of young males (aged 15-24) and single
people were used because of their increased likelihood of victimization under the routine
activity framework (Cohen & Felson, 1979; Kennedy & Forde, 1990). Young males have
also been associated with increased criminal activity (Hirschi & Gottfredson, 1983).
At this point the modifiable areal unit problem (MAUP) should be acknowledged.
The MAUP is an issue for all spatially-referenced data and refers to the aggregation of
data from individuals to larger units of analysis (Andresen & Malleson, 2013). For the
current study, this means aggregating individual-level data (e.g. visible minority status)
to the dissemination area level. Dissemination areas do not represent natural units; they
are defined somewhat arbitrarily by population. Defining areal units in different ways,
either by changing the spatial scale or by shifting their boundaries could result in
different findings (Andresen, 2014). Indeed, Andresen and Malleson (2013) found
important differences between smaller and larger areal units (dissemination areas and
census tracts) in their spatial regressions. To truly address the MAUP multiple spatial
scales of analysis should be used. Because the current study only uses dissemination
areas, it is possible that the findings would not hold at, for example, the census tract
level.7 Nevertheless, the choice of spatial unit has also been found to have little effect on
substantive results (Wooldredge, 2002).
Another issue relating to the census data used in this study is worth mentioning
briefly. Vancouver has a large homeless population; a 2018 count found 2,181 homeless
people living in the city (“More than half of Vancouver’s homeless population”, 2018).
Census data are gathered through a questionnaire sent to Canadian households.
Statistics Canada (2009) does note that, when possible, the short form of the census is
administered to homeless individuals in shelters. Still, it could be that data from an
important segment of the population is being missed. Moreover, given the uneven
7 To avoid the ecological and atomistic fallacies, all inference is conducted at the dissemination area level. See Andresen (2014) for a discussion of this issues.
28
distribution of homeless individuals across Vancouver, it is likely that certain
dissemination areas are affected more than others. Unfortunately, there is little that
could reasonably be done to address this issue.
4.2. Methods
To create a single dataset, the above three data sources were merged using a
spatial join function in ArcMap 10.3. Of the seven property crimes from the VPD data,
two were excluded from this analysis: commercial and residential break-and-enter. The
population at risk for these two crime types has less to do with the ambient population of
a given area than the other five. More accurate crime rate denominators for commercial
and residential break-and-enter might be the number of commercial units and
households in a given area, respectively. Mischief and other thefts may involve persons
(e.g. purse snatching) and people moving through their environment as part of their
routine activities often use vehicles or bicycles.
To permit comparison, crime rates with both residential and ambient population
denominators were constructed. These ten rate variables were then used as dependent
variables in ten separate regression models. Because there were different patterns of
spatial autocorrelation in the dependent and independent variables used in this study, a
spatial error model, as opposed to a lag model, was used. The spatial error models were
initially run using the spatial data analysis software GeoDa. GeoDa is available for free
download through the University of Chicago (https://spatial.uchicago.edu/software).
Proper Queen’s contiguity orders for the models were determined with Moran’s I
significance testing of the error residuals. Only the mischief residential rate model
required second-order Queen’s contiguity to filter out the spatial autocorrelation; all the
others tested insignificant (p > 0.05).
When the models were run in GeoDa, all ten were significant on the spatial
dependence test (p < 0.05), justifying the use of a spatial regression technique.
However, all ten models were also significant on the Breusch-Pagan test, indicating
heteroskedasticity in the data. Because of this finding, it was necessary to use software
that controls for heteroskedasticity along with spatial autocorrelation: GeoDaSpace
(https://spatial.uchicago.edu/software). Spatial error models were once again run using a
GMM estimate with KP HET standard errors. Full models with all 17 independent
29
variables were produced for both the resident and ambient population-based rates for
each of the five crime types. In terms of specification for final models, the least
significant variables were removed first, then the regressions were re-run. This process
was repeated until all the remaining variables were significant at the p < 0.10 level
(Andresen, 2006b). This was done to minimize the chances of omitted variable bias in
the final models. These full and final models permit comparison between residential and
ambient population-based crime rates on model fit, variable retention, and significant
relationships.
30
Chapter 5. Results
5.1. Descriptive Statistics, Dependent Variables
Table 5.1. Descriptive statistics for dependent variables
Minimum Maximum Mean Standard Deviation
Mischief (residential) 0 117.55 5.815 8.093
Mischief (ambient) 0 122.13 8.952 14.312
Theft from vehicle (residential) 0 111.111 14.002 12.734
Theft from vehicle (ambient) 0 293.112 24.174 34.498
Theft of vehicle (residential) 0 15.564 2.103 2.38
Theft of vehicle (ambient) 0 73.278 3.91 7.536
Theft of bicycle (residential) 0 96.026 3.306 6.596
Theft of bicycle (ambient) 0 146.556 4.73 11.218
Other theft (residential) 0 471.287 6.338 32.615
Other theft (ambient) 0 219.814 3.644 13.413
Note: All rates are per 1,000 persons, n = 978
Table 5.1. presents descriptive statistics for the dependent variables used in the
current study. The crime rates range from zero (all ten rates) to a maximum of 471.287
(other theft, residential). Overall, the results suggest that there is something different
about the other theft crime type. The ranges, means, and standard deviations for
ambient population-based rates of mischief, theft of vehicle, theft from vehicle, and theft
of bicycle are consistently greater, compared to their residential counterparts. The
reverse is true for the other theft crime type.
Maps depicting the spatial patterning of each of the dependent variables, created
using ArcMap 10.3, are now presented and discussed. Overall, the visual differences in
the spatial patterns between the residential and ambient population-based maps are
striking, and reinforce the importance of considering this measure.
31
Figure 5.1. Residential population-based rates of mischief
Figure 5.2. Ambient population-based rates of mischief
32
Figure 5.3. Residential population-based rates of theft from vehicle
Figure 5.4. Ambient population-based rates of theft from vehicle
33
Figure 5.5. Residential population-based rates of theft of vehicle
Figure 5.6. Ambient population-based rates of theft of vehicle
34
Figure 5.7. Residential population-based rates of theft of bicycle
Figure 5.8. Ambient population-based rates of theft of bicycle
35
Figure 5.9. Residential population-based rates of other theft
Figure 5.10. Ambient population-based rates of other theft
36
Compared to residential population-based property crime rates, nearly all of the
above maps demonstrate lower ambient rates in Vancouver’s north-central downtown
area. Because of the larger ambient population in these dissemination areas (see
Figures 4.1., 4.2., & 4.3.), crime rates using this denominator will necessarily be lower.
Theft of bicycle (Figures 5.7. & 5.8) provides a clear illustration of this trend. Immediately
noticeable is the lowered ambient population-based rate in Vancouver’s busy Stanley
Park (the northernmost dissemination area), that attracts many people engaged in their
routine activities. The spatial patterns do not appear to change much for rates of other
theft (Figures 5.9. & 5.10.), depending on the population denominator used. However, it
is apparent that several of the dissemination areas with higher residential population-
based rates have lower rates of other theft when the ambient population is used.
Another noticeable trend concerns the identification of new, often more
dispersed, hot spots. The maps for theft of vehicle (Figures 5.5. & 5.6.), for example,
depict hot spots in the eastern and southern parts of the city, when the ambient
population is used. Similarly, the highest ambient rate dissemination areas for mischief
(Figure 5.2.) are far more dispersed, compared to the residential population-based map
(Figure 5.1.). These types of findings may have implications for crime prevention
initiatives and police operations.
A final trend worth noting concerns a specific dissemination area in Vancouver’s
northeastern corner. Hastings Park located in this dissemination area, contains a
popular summer fair, an amusement park, a horse track and an arena. These venues
attract many visitors from Vancouver and its surrounding municipalities. The parking lots
that surround the park are known by locals to be hotspots for both forms of auto theft.
Indeed, the maps for residential population-based rates of theft from and theft of vehicle
(Figures 5.3. & 5.5.) show moderate levels of these crime types in this area.
Interestingly, this dissemination area falls into the lowest rate category for both crime
types when the ambient population is used (Figures 5.4. & 5.6.). When the large crowds
that the fair attracts are accounted for with the ambient population measure, it becomes
apparent that rates of both forms of auto theft may not actually be disproportionately
high around Hastings Park.
37
5.2. Descriptive Statistics and Correlations, Independent Variables
Table 5.2. Descriptive statistics for independent variables
Minimum Maximum Mean Standard Deviation
Aboriginal (%) 0 40.404 2.259 3.561
Ethnic heterogeneity 0 93.378 39.551 16.367
Visible minorities (%) 7.407 100 50.956 25.355
Immigrants (%) 9.532 88.034 41.85 14.934
Recent immigrants (%) 0 27.933 5.78 4.037
Moved within 1 year (%) 0 62.793 16.225 7.71
Single persons (%) 23.81 90.121 43.411 9.431
Lone parents (%) 0 62.791 16.159 7.085
Males aged 15-24 (%) 0 13.787 5.884 2.421
Unemployed (%) 0 30.303 5.771 3.345
Receiving government assistance (%) 1.1 67.9 9.325 6.042
Low income designation (%) 4.213 78.306 17.985 8.175
Median income (thousands, CAD) 11.504 68.736 33.742 9.701
Subsidized housing (%) 0 90.9 7.864 16.833
Houses under major repair (%) 0 41.159 6.423 4.739
Median income (thousands, CAD), X13 1 -.118** 0.047 -0.048 .725**
Subsidized housing (%), X14 1 .224** .450** 0.023
Houses under major repair (%), X15 1 .259** .106**
Rented households (%), X16 1 .269**
Post-secondary education (%), X17 1
* p < 0.05, ** p < 0.01
40
Only two of the significant relationships are above the commonly used 0.8
threshold for multicollinearity. The relationship between visible minorities and immigrants
(ρ = 0.906, p < 0.01) is hardly surprising; in Vancouver/Canada immigrants are often
also visible minorities (Chard & Renaud, 1999; Ley & Smith, 2000). It would therefore
make sense that dissemination areas with greater percentages of immigrants would
have greater percentages of visible minorities, and vice versa. Those receiving
government assistance and median income are also highly correlated (ρ = -0.834, p <
0.01). This relationship is intuitive: as the percentage of those receiving government
assistance in a dissemination area increases, median incomes decrease. All four of the
above variables were kept in this analysis to avoid omitted variable bias.
None of the stronger relationships (ρ > 0.5 or ρ < -0.5) between the variables
used in this study are particularly surprising. Males aged 15-24 are positively associated
with visible minorities (ρ = 0.605, p < 0.01) and immigrants (ρ = 0.553, p < 0.01). This
finding simply speaks to demographic trends in Vancouver dissemination areas. Single
persons have a positive relationship with rented households (ρ = 0.733, p < 0.01). Since
single people do not have a partner to help purchase a home, it is unsurprising that
dissemination areas with a greater percentage of single people also have a greater
percentage of rented households, particularly given Vancouver’s housing prices (Lee,
2017).
Government assistance was found to be positively associated with visible
minorities (ρ = 0.500, p < 0.01) and lone parents (ρ = 0.560, p < 0.01). It makes sense
that in dissemination areas with higher percentages of lone parents a greater proportion
of the population relies on government assistance, since single parents only have one
income to support their children. Government assistance has a negative relationship with
post-secondary education (ρ = -0.705, p < 0.01). Negative relationships also exist
between post-secondary education and both visible minorities (ρ = -0.602, p < 0.01) and
lone parents (ρ = -0.541, p < 0.01). As percentages of visible minorities and lone parents
in a dissemination area increase, the percentage of residents with post-secondary
education decreases.
Median income has associations, both positive and negative, with several
variables. Vancouver dissemination areas with higher percentages of post-secondary
education typically have higher median incomes (ρ = 0.725, p < 0.01). Higher median
41
income levels are also associated with lower percentages of immigrants (ρ = -0.600, p <
0.01) and visible minorities (ρ = -0.663, p <0.01). Lastly, as the percentage of lone
parents in Vancouver dissemination areas increases, median incomes decrease (ρ = -
0.557, p < 0.01).
Some weaker relationships were somewhat unexpected. For instance, there are
significant negative relationships between ethnic heterogeneity and visible minorities (ρ
= -0.195, p < 0.01), immigrants (ρ = -0.237, p < 0.01), and recent immigrants (ρ = -0.156,
p < 0.01). It may be that these groups cluster, meaning that that there is less overall
heterogeneity in these dissemination areas. An interesting finding is that rented
households have negative relationships with visible minorities (ρ = -0.400, p < -0.02) and
immigrants (ρ = -0.372, p < 0.01). In their 2001 study, Ley and Murphy found little
difference in rental affordability stress between immigrants and non-immigrants, as well
as between visible minorities and “the rest (white)” (p. 143). Ley and Murphy (2001)
point out that Vancouver’s immigrant population is different from the rest of Canada, with
many immigrants coming from Hong Kong and Taiwan for business. These immigrants
(and visible minorities) are likely wealthier, and more likely to purchase a home,
explaining these dissemination area level trends.
5.3. Multivariate Results
Pseudo R2 measures the correlation between the actual dependent variables and
their predicted values as a measure for goodness-of-fit. Overall, Pseudo R2 values were
low across the regression models for the five crime types. Across the five crime types,
Pseudo R2 values were consistently higher in both the full and final models for the
residential population-based crime rates. Models ranged from having two to six
significant independent variables. Each of the seventeen independent variables used in
this study were significant in at least one of the models. There was no switching of
coefficient signs for significant variables from the full to final models, indicating that the
‘qualitative’ results remained the same and that the removal of statistically insignificant
variables did not lead to omitted variable bias. For the 10 different crime rate types, there
were 35 instances of variable retention from the full to final models. Eight variables that
were initially insignificant became significant in the final models. Five variables became
insignificant in the final models. A finding like this usually suggests multicollinearity.
42
However, all five of these independent variables were of marginal significance in the full
models (0.1 < p < 0.05), indicating that this is likely not all that serious of a problem.
5.3.1. Mischief
The Pseudo R2 values for the full and final models for the residential population-
based mischief rate are 0.217 and 0.174, respectively. The percentage of single persons
in Vancouver dissemination areas was identified as the biggest predictor of mischief
across both models (Full residential: β = 0.265, p < 0.01; Final residential: β = 0.28, p <
0.01). The percentage of rented households is negatively associated with rates mischief
across both models (Full residential: β = -0.051, p < 0.05; Final residential: β = -0.028, p
< 0.1). The percentages of those receiving government assistance (β = -0.155, p < 0.05)
and those with post-secondary education (β = -0.067, p < 0.1) are both negatively
associated with rates of mischief in the final residential model.
For the full ambient population model of mischief, the Pseudo R2 value is 0.116,
while the final model has a value of 0.105. Six variables have significant relationships
with ambient population-based rates of mischief, and these variables were all retained in
the final models. The percentage of aboriginals in a dissemination area emerged as the
most important predictor of mischief (Full ambient: β = 0.719, p < 0.01; Final ambient: β
= 0.87, p < 0.01). Increased percentages of those receiving government assistance (Full
ambient: β = 0.401, p < 0.05; Final ambient: β = 0.37, p < 0.05), low income designation
(Full ambient: β = 0.23, p < 0.05; Final ambient: β = 0.165, p < 0.1), and those with post-
secondary education (Full ambient: β = 0.141, p < 0.05; Final ambient: β = 0.151, p <
0.01) are all associated with higher ambient population-based rates of mischief. The
percentages of both residents who moved into a dissemination area within the last year
(Full ambient: β = -0.184, p < 0.01; Final ambient: β = -0.187, p < 0.05) and lone parents
(Full ambient: β = -0.224, p < 0.05; Final ambient: β = -0.245, p < 0.05) are negatively
associated with mischief.
43
The regression results for mischief are presented in Table 5.4. below:
Table 5.4. Spatial regression results for mischief
Full residential model
Final residential model
Full ambient model
Final ambient model
Aboriginal (%) 0.003 0.719*** 0.87***
Ethnic heterogeneity 0.003 -0.048
Visible minorities (%) -0.01 0.002
Immigrants (%) -0.042 -0.17**
Recent immigrants (%) 0.016 0.191
Moved within 1 year (%) 0.071 -0.184*** -0.187**
Single persons (%) 0.265*** 0.28*** -0.068
Lone parents (%) -0.085 -0.224** -0.245**
Males aged 15-24 (%) 0.258 -0.049
Unemployed (%) -0.036 -0.037
Receiving government assistance (%)
-0.123 -0.155** 0.401** 0.37**
Low income designation (%)
0.065 0.23** 0.165*
Median income (thousands, CAD)
-0.029 -0.094
Subsidized housing (%) 0.02 0.015
Houses under major repair (%)
0.044 0.028
Rented households (%) -0.051** -0.028* -0.038
Post-secondary education (%)
-0.066 -0.067* 0.141** 0.151***
Pseudo R2 0.217 0.174 0.116 0.105
n = 978, * p < 0.1, ** p < 0.05, *** p < 0.01
44
5.3.2. Theft from Vehicle
The full and final models for residential population-based rates of theft from
vehicle have respective Pseudo R2 values of 0.168 and 0.145. As percentages of visible
minorities (Full residential: β = -0.164, p < 0.01; Final residential: β = -0.15, p < 0.01) and
lone parents (Full residential: β = -0.181, p < 0.1; Final residential: β = -0.238, p < 0.05)
increase, rates of theft from vehicle decrease. As the median income of a dissemination
area increases, rates of theft from vehicle also decrease (Full residential: β = -0.206, p <
0.05; Final residential: β = -0.229, p < 0.05). In the full model, the percentage of
subsidized housing has a positive relationship with residential population-based rates of
theft from vehicle (β = 0.076, p < 0.1), while the percentage of those receiving
government assistance has a negative one (β = -0.264, p < 0.05). Lastly, houses under
major repair emerged as a significant predictor of theft from vehicle in the final
residential model (β = 0.139, p < 0.1). In other words, as the percentage of houses under
major repair in a dissemination area increases, so does the residential population-based
rate of theft from vehicle.
The Pseudo R2 values for the full and final ambient population-based models of
theft from vehicle are 0.075 and 0.054, respectively. All four of the significant
associations were maintained across the full and final models. The percentages of those
receiving government assistance (Full ambient: β = 0.603, p < 0.1; Final ambient: β =
1.166, p < 0.01) and those with post-secondary education (Full ambient: β = 0.258, p <
0.1; Final ambient: β = 0.385, p < 0.01) are associated with increased ambient
population-based rates of theft from vehicle. Those receiving government assistance is
also the most important predictor for both the full and final ambient models. As the
percentages of single persons (Full ambient: β = -0.532, p < 0.01; Final ambient: β = -
0.406, p < 0.05) and lone parents (Full ambient: β = -0.48, p < 0.05; Final ambient: β = -
0.374, p < 0.1) increase, rates of theft from vehicle decrease.
45
The regression results for theft from vehicle are presented in Table 5.5. below:
Table 5.5. Spatial regression results for theft from vehicle
Full residential model
Final residential model
Full ambient model
Final ambient model
Aboriginal (%) -0.127 0.401
Ethnic heterogeneity -0.048 -0.154
Visible minorities (%) -0.164*** -0.15*** -0.175
Immigrants (%) 0.029 -0.123
Recent immigrants (%) 0.083 0.136
Moved within 1 year (%) 0.068 -0.239
Single persons (%) 0.151 -0.532*** -0.406**
Lone parents (%) -0.181* -0.238** -0.48** -0.374*
Males aged 15-24 (%) -0.3 -0.695
Unemployed (%) -0.043 0.199
Receiving government assistance (%)
-0.264* 0.603* 1.166***
Low income designation (%)
-0.004 0.249
Median income (thousands, CAD)
-0.206** -0.229** -0.134
Subsidized housing (%) 0.076* 0.047
Houses under major repair (%)
0.128 0.139* 0.077
Rented households (%) -0.021 0.054
Post-secondary education (%)
-0.089 0.258* 0.385***
Pseudo R2 0.168 0.145 0.075 0.054
n = 978, * p < 0.1, ** p < 0.05, *** p < 0.01
46
5.3.3. Theft of Vehicle
The Pseudo R2 values for the full and final models for the residential population-
based rate of theft of vehicle are 0.105 and 0.091, respectively. The biggest predictor
across both models is single persons (Full residential: β = 0.031, p < 0.05; Final
residential: β = 0.042, p < 0.01). This finding means that as the percentage of single
persons in a dissemination area increases, so does the rate of theft of vehicle. The
percentages of immigrants (Full residential: β = -0.024, p < 0.1; Final residential: β = -
0.024, p < 0.01) and those with post-secondary education (Full residential: β = -0.025, p
< 0.05; Final residential: β = -0.026, p < 0.01) have negative relationships with the rate of
theft of vehicle. The percentage of houses under major repair in a dissemination area
became significant in the final model for theft of vehicle (β = 0.031, p < 0.1). Ethnic
heterogeneity also emerged as a significant negative predictor of residential population-
based rates of theft of vehicle in the final model (β = -0.012, p < 0.05).
For the ambient population-based rates of theft of vehicle, the Pseudo R2 value
for the full model is 0.057, while that of the final model is 0.045. For this crime rate, all
three significant variables were retained in the final model. The largest predictor of theft
of vehicle was found to be the percentage of aboriginals (Full ambient: β = 0.249, p <
0.05; Final ambient: β = 0.261, p < 0.01). The percentage of people on government
assistance (Full ambient: β = 0.181, p < 0.05) is also associated with increased rates of
theft of vehicle. Dissemination areas with a greater percentage of lone parents are
associated with lower ambient population-based rates of theft of vehicle (Full ambient: β
= -0.094, p < 0.1; Final ambient: β = -0.115, p < 0.05).
47
The regression results for theft of vehicle are presented in Table 5.6. below:
Table 5.6. Spatial regression results for theft of vehicle
Full residential model
Final residential model
Full ambient model
Final ambient model
Aboriginal (%) 0.05 0.249** 0.261***
Ethnic heterogeneity -0.008 -0.012** -0.026
Visible minorities (%) 0.002 -0.002
Immigrants (%) -0.024* -0.024*** -0.042
Recent immigrants (%) 0.014 0.008
Moved within 1 year (%) 0.007 -0.049
Single persons (%) 0.031** 0.042*** -0.026
Lone parents (%) -0.01 -0.094* -0.115**
Males aged 15-24 (%) -0.057 0.006
Unemployed (%) -0.033 -0.035
Receiving government assistance (%)
0.002 0.181** 0.148**
Low income designation (%)
-0.01 -0.035
Median income (thousands, CAD)
-0.018 -0.034
Subsidized housing (%) -0.001 0.01
Houses under major repair (%)
0.026 0.031* -0.008
Rented households (%) 0.002 0.006
Post-secondary education (%)
-0.025** -0.026*** 0.026
Pseudo R2 0.105 0.091 0.057 0.045
n = 978, * p < 0.1, ** p < 0.05, *** p < 0.01
48
5.3.4. Theft of Bicycle
The respective Pseudo R2 values for the full and final residential population-
based rate models for theft of bicycle are 0.196 and 0.173. The biggest predictor across
both models is those receiving government assistance (Full model: β = -0.203, p < 0.01;
Final model: β = -0.217, p < 0.01). As the percentage of residents receiving government
assistance in a dissemination area increases the residential population-based rate of
theft of bicycle decreases. The percentage of immigrants is also negatively associated
with the rate of theft of bicycle (Full residential: β = -0.084, p < 0.05; Final residential: β =
-0.07, p < 0.01). For both models, the percentages of single persons (Full residential: β =
0.156, p < 0.01; Final residential: β = 0.152, p < 0.01) and low income designation (Full
residential: β = 0.108, p < 0.1; Final residential: β = 0.115, p < 0.05) are positively
associated with theft of bicycle. When the percentages of each of these variables
increases, so does the residential population-based rate of theft of bicycle in Vancouver
dissemination areas.
For the ambient population-based models of theft of bicycle in Vancouver, the full
model has a Pseudo R2 value of 0.121, while the final model has a value of 0.11. This
time, the percentage of males aged 15-24 emerged as the largest significant predictor
across both full and final models (Full ambient: β = -0.664, p < 0.05; Final ambient: β = -
0.593, p < 0.01). Dissemination areas with a greater percentage of young males have
lower ambient population-based rates of theft of bicycle. Ethnic heterogeneity (Full
model: β = -0.044, p < 0.1; Final model: β = -0.052, p < 0.05) and visible minorities (Full
model: β = -0.059, p < 0.1; Final model: β = -0.099, p < 0.01) were also found to be
negatively associated with theft of bicycle. Low income is a positive predictor across
both models (Full model: β = 0.156, p < 0.05; Final model: β = 0.179; p < 0.01). Finally,
the percentage of those with post-secondary education are associated with increased
theft of bicycle (β = 0.062, p < 0.1), but only in the full ambient model.
49
The regression results for theft of bicycle are presented in Table 5.7. below:
Table 5.7. Spatial regression results for theft of bicycle
Full residential model
Final residential model
Full ambient model
Final ambient model
Aboriginal (%) 0.048 0.149
Ethnic heterogeneity -0.013 -0.044* -0.052**
Visible minorities (%) 0.016 -0.059* -0.099***
Immigrants (%) -0.084** -0.07*** -0.07
Recent immigrants (%) -0.032 0.02
Moved within 1 year (%) 0.058 -0.027
Single persons (%) 0.156*** 0.152*** -0.085
Lone parents (%) -0.062 0.02
Males aged 15-24 (%) -0.133 -0.664** -0.593***
Unemployed (%) 0.03 0.102
Receiving government assistance (%)
-0.203*** -0.217*** -0.075
Low income designation (%)
0.108* 0.115** 0.156** 0.179***
Median income (thousands, CAD)
0.003 -0.08
Subsidized housing (%) 0.031 0.008
Houses under major repair (%)
-0.029 -0.099
Rented households (%) -0.018 0.016
Post-secondary education (%)
-0.008 0.062*
Pseudo R2 0.196 0.173 0.121 0.11
n = 978, * p < 0.1, ** p < 0.05, *** p < 0.01
50
5.3.5. Other Theft
For the final crime type examined, the full and final residential population-based
models for other theft have Pseudo R2 values of 0.051 and 0.041, respectively. There is
not consistency across the full and final models in terms of the most important predictor.
Both ethnic heterogeneity (Full residential: β = 0.138, p < 0.1; Final residential: β =
0.118, p < 0.1) and the percentage of single persons (Full residential: β = 0.61, p < 0.01;
Final residential: β = 0.671, p < 0.01) are positively associated with rates of other theft.
In only the full residential model, median income has a positive relationship with other
theft rates (β = 0.322, p < 0.1), while the unemployment percentage has a negative one
(β = -0.663, p < 0.1). In the final residential model, both the percentages of immigrants
(β = 0.202, p < 0.1) and those receiving government assistance (β = -0.527, p < 0.05)
emerged as significant predictors.
The full ambient population-based for other theft has a Pseudo R2 value of 0.047,
while the final model has a value of 0.04. Compared to the residential population-based
models, the ambient ones retain more variables. Across both models, the percentage of
young males was found to be the biggest predictor of other theft (Full ambient: β = -
0.435, p < 0.05; Final ambient: β = -0.549, p < 0.05). This finding means that as the
percentage of young males in a dissemination area increases, the ambient population-
based rates of other theft actually decrease. This somewhat counterintuitive finding will
be discussed more in the following section. Government assistance also has a negative
relationship with ambient population-based rates of other theft (Full ambient: β = -0.264,
p < 0.05; Final ambient: β = -0.247, p < 0.01). The percentages of recent immigrants
(Full ambient: β = 0.227, p < 0.1; Final ambient: β = 0.296, p < 0.05) and single persons
(Full ambient: β = 0.13, p < 0.1; Final ambient: β = 0.188, p < 0.01) have positive
relationships with the dependent variable. In the final model, the percentage of lone
parents is a significant predictor of increased ambient population-based rates of other
theft (β = 0.186, p < 0.05).
51
The regression results for other theft are presented in Table 5.8. below:
Table 5.8. Spatial regression results for other theft
Full residential model
Final residential model
Full ambient model
Final ambient model
Aboriginal (%) 0.317 0.263
Ethnic heterogeneity 0.138* 0.118* 0.043
Visible minorities (%) 0.056 0.006
Immigrants (%) 0.145 0.202* 0.066
Recent immigrants (%) 0.347 0.227* 0.296**
Moved within 1 year (%) 0.153 -0.053
Single persons (%) 0.61*** 0.671*** 0.13* 0.188***
Lone parents (%) -0.106 0.112 0.186**
Males aged 15-24 (%) 0.552 -0.435** -0.549**
Unemployed (%) -0.663* -0.017
Receiving government assistance (%)
-0.267 -0.527** -0.264** -0.247***
Low income designation (%)
0.275 0.082
Median income (thousands, CAD)
0.322* 0.107
Subsidized housing (%) -0.047 -0.019
Houses under major repair (%)
0.075 0.014
Rented households (%) 0.026 0.045
Post-secondary education (%)
-0.048 -0.007
Pseudo R2 0.051 0.041 0.047 0.04
n = 978, * p < 0.1, ** p < 0.05, *** p < 0.01
52
Chapter 6. Discussion and Conclusions
6.1. Spatial Findings
As discussed above, the differences in spatial patterns of rates of mischief, theft
from vehicle, theft of vehicle, theft of bicycle, and other theft are often striking, depending
on the population denominator used. The finding that hotspots in Vancouver’s downtown
area decrease in intensity when the ambient population is used speaks to Boggs’ (1965)
assertion regarding spuriously high crime rates in central business districts. When a
more appropriate population at risk is used, the risk of being a victim of property crime is
not substantially higher in downtown Vancouver dissemination areas. This finding is
consistent with the later work of Andresen (2011) and Mburu and Helbich (2016) on
aggregate violent crime. It should be noted, that Andresen’s (2011) work only
demonstrated this finding at the dissemination area level; when spatial patterns of violent
crime were examined at the census tract level, the same trends did not hold.
New, and often more dispersed, clusters of high crime rate dissemination areas
were also identified in the current study when maps of ambient population-based rates
were compared to residential ones. These results echo the work of Malleson and
Andresen (2015b, 2016), who also identified new statistically significant clusters of
aggregate violent crime and theft from persons offenses when an ambient population
measure was used. Overall, these findings underscore the importance of considering the
population at risk. Because such a different picture of environmental risk is painted when
the ambient population is used as the crime rate denominator, these findings seriously
bring into question the near-ubiquitous use of the residential population.
The findings from the current study also have implications from a crime
prevention policy perspective.8 Police and policymakers rely on accurate measures of
environmental risk to answer questions such as:
8 See Andresen and Jenion (2008) for a more in-depth discussion of the ambient population in relation to crime prevention at the primary, secondary, and tertiary levels.
53
• Where should police direct their patrols in response to increased rates of other theft across the city?
• Which areas would benefit most from school outreach programs?
• Which parking lots should receive closed-circuit television cameras as a theft of and theft from vehicle crime prevention measure?
These, and other such policy decisions depend on accurate measures of environmental
risk for particular crime types. The ambient population provides a different lens for
assessing risk, as clearly demonstrated by the maps presented in the results section.
6.2. Inferential Findings
The multivariate results are the primary focus of the current study. Figure 6.1.,
presented below, is a summary table, to permit easier comparison of significant
relationships, variable retention, and Pseudo R2 values between regression models. A
quick glance at Figure 6.1. reveals important differences between regression models for
disaggregated property crime rates that use either residential or ambient population
denominators.
54
Figure 6.1. Regression summary table
55
As discussed in the literature review, only three prior studies have included
ambient population-based rates as dependent variables in regression models. As such,
much of the current discussion will compare the current study’s findings to the work of
Andresen (2006b, 2011) and Andresen and Brantingham (2007). As mentioned, the
current study differentiates itself from their work in four important ways. First, this study
was conducted at a finer spatial scale, the dissemination area. Only Andresen’s (2011)
study examined dissemination areas; the other two studies were conducted at the
census tract level. Second, the present work examines five disaggregated property
crime types. The previous three studies all used aggregate measures of crime, such as
total violent crime (Andresen, 2011) or automotive theft (Andresen, 2006b), consisting of
both theft of and theft from vehicle. It may be that aggregation masks important trends
and/or relationships that only become apparent when disaggregated crime types are
used. The third way in which the current study differentiates itself from earlier works
concerns the use of relatively current crime and census data. Andresen (2006b) made
use of data from 1996, while both Andresen and Brantingham (2007) and Andresen
(2011) used data from 2001. While these studies examined Vancouver as well, a lot may
have changed in the past 15-20 years. Lastly, the current study makes use of a different,
novel ambient population measure. The work of Andresen (2006b, 2011) and Andresen
and Brantingham (2007) all used 24-hour average population estimates from LandScan
Global Population Database.
The current study’s finding that Pseudo R2 values are lower for full and final
models for ambient population-based crime rates, compared to their residential
counterparts, is somewhat unexpected. Previous studies have consistently found the
opposite (Andresen, 2006b, 2011; Andresen & Brantingham, 2007). Only for aggregate
violent crime did Andresen (2006b) find a higher Pseudo R2 value for the ambient model.
The opposite trend of the current study to prior works may be explained by the
differences between the studies detailed above. The generally low Pseudo R2 values
suggest that there is more to explain in the spatial patterns of crime at the dissemination
area level than the theoretically-informed census variables permit.
In terms of variable retention across full and final models, there is either more or
equal retention for ambient population-based rates. Final models for ambient population-
based rates also typically have either a greater or equivalent number of significant
56
variables, compared to final residential models. The only exception is theft of vehicle.
These findings show the limited value of using Pseudo R2 values for model assessment.
For the three prior inferential studies that used the ambient population as a crime rate
denominator, there was no consistent pattern in terms of the number of significant
variables.
Only a handful of significant variables are consistent across final residential and
ambient models in the direction of their relationship with the particular crime type. The
percentage of lone parents in a dissemination area has a negative relationship with theft
from vehicle, regardless of the population denominator used. Whichever population at
risk is accounted for, greater percentages of lone parents are associated with a
decrease in this crime type. This relationship may be a question of suitable targets; lone
parents may be less able to afford a vehicle to be broken into.
Low income has a consistent positive relationship with theft of bicycle for both
residential and ambient population-based rate models. This finding may speak to social
disorganization processes regarding low socioeconomic status and crime (Sampson &
Groves, 1989). For other theft, the percentage of single persons in a dissemination area
is a positive predictor for both residential and ambient models. The routine activities of
this population segment may bring them out of the home more, which may create more
opportunities for victimization (Cohen & Felson, 1979). Finally, government assistance is
negatively associated with both residential and ambient population-based rates of other
theft. Andresen and Brantingham (2007) found a similar consistent negative relationship
for percentages of those receiving government assistance across residential and
ambient models for aggregate property crime at the census tract level. In this case,
routine activity theory may provide an explanation for this relationship. People receiving
government assistance can be assumed to have low socioeconomic status, and may not
possess many goods that would be considered suitable targets.
Immediately apparent from Figure 6.1. are the differences between the final
residential and ambient population-based rate models. Many variables that are
significant predictors in one model are insignificant in the other. For example, low
income designation is a positive predictor of ambient population-based rates of mischief.
This finding speaks to social disorganization theory, and the link between low
socioeconomic status and crime (Sampson & Groves, 1989). However, the same
57
relationship does not hold up when the residential population is used. From a theoretical
perspective, the finding is somewhat troubling. If a relationship predicted by social
disorganization theory holds only when the ambient population is used, it brings into
question the exclusive use of the residential population as a crime rate denominator for
theory testing.
In the context of routine activity theory, the relationships between the percentage
of single persons in a dissemination area and rates of mischief, theft of vehicle, and theft
of bicycle are quite interesting. For residential population-based rates of these three
crime types, the percentage of single persons is a significant positive predictor. This
relationship is consistent with Cohen and Felson’s (1979) findings. Single people’s
routine activities take them out of the home more often and put them at greater risk of
criminal victimization (Cohen & Felson, 1979). Yet, when the ambient population is used
all three of these relationships become insignificant. Both this and the above finding
highlight the impact an alternative denominator can have on theoretically-predicted
relationships.
The differences between final residential and ambient population-based rate
models are also important when it comes to policy-relevant variables. While ethnic
heterogeneity or the percentage of lone parents in a dissemination area cannot
(reasonably) be controlled, policies enacted by various levels of government on
subsidized housing, for instance, can affect crime rates. In the current study,
percentages of post-secondary education are negatively associated with residential
population-based rates of theft of vehicle. Policymakers might think that improving
access to post-secondary education could have long-term effects on rates of theft of
vehicle. When the ambient population is used, however, this relationship disappears. If
this measure provides a more accurate indication of environmental risk than the
residential population, policies enacted to increase post-secondary education may be
ineffective. A similar trend exists for those receiving government assistance and rates of
theft of bicycle. When the residential population is used, the relationship is negative, but
becomes insignificant in the ambient model. Policy decisions depend on accurate
assessments of the relationships between crime risk and sociodemographic and
socioeconomic indicators. These findings suggest that alternative population measures
should be considered alongside the residential population, when conducting research
relevant to policy.
58
While most of the differences between final residential and ambient population-
based rate models involve variables falling in and out of significance, there are two9
instances of relationships switching direction between final models for residential and
ambient population-based crime rates. The percentage of those receiving government
assistance has a negative relationship with residential population-based rates of
mischief, yet the relationship is positive for the ambient rate. This finding means that
when the residential population of a dissemination area is controlled for, increased
percentages of those receiving government assistance are associated with lower rates
of mischief, and vice versa. However, when it is the ambient population that is controlled
for, both percentages of those receiving government assistance and rates of mischief
vary together. Interestingly, when the ambient population is used, the relationship
conforms to social disorganization expectations regarding low socioeconomic status. Yet
the negative relationship in the residential model may speak more to routine activity
theory. In dissemination areas with lower percentages of those receiving government
assistance, there may be more suitable targets for mischief.
A similar trend exists for postsecondary education and mischief. Postsecondary
education is negatively associated with residential population-based rates of mischief,
but the relationship is positive when the ambient population is controlled for instead.
These findings are particularly important, because they demonstrate the impact the use
of a theoretically-informed alternative denominator can have on results. When the
number of people that visit an area are considered, as opposed to the number of people
who sleep in that area, significant relationships can switch direction. Worth noting, is that
there was no switching of signs for socioeconomic or sociodemographic variables in the
studies conducted by Andresen (2006b, 2011) and Andresen and Brantingham (2007).
From a policy perspective, these findings are perhaps even more worrisome than
variables going in and out of significance between final residential and ambient
population-based crime rate models. Crime reduction policies are often informed by
relationships between residential population-based crime rates and sociodemographic
and socioeconomic indicators. If more accurate population measures (i.e. the ambient
9 There was one other instance of a relationship switching direction between models for residential and ambient population-based crime rates: the percentage of those receiving government assistance for the theft from vehicle crime type. However, this relationship was not significant in the final residential population-based model.
59
population) suggest that these same relationships are in the opposite direction, these
policies could potentially increase crime. Taken together, the findings from this study
indicate that the use of a theoretically-informed crime rate denominator impacts results
in a substantial way. There are important differences in spatial patterns, Pseudo R2
values, variable retention, and trends in significant relationships between crime rates
using either residential or ambient population denominators. As discussed, there are
differences between the current study and the work of Andresen (2006b, 2011) and
Andresen and Brantingham (2007) in both design and results. Nevertheless, the overall
story told is the same. Clearly, the question of the most appropriate crime rate
denominator is not just an obscure measurement issue to be acknowledged in passing;
the population at risk matters.
6.3. Limitations
Several limitations of the current study have already been discussed, including
the MAUP and the use of a single spatial scale, the dark figure of crime, and issues
pertaining to the representation of homeless and poorer populations in all three data
sources used in this study. The ambient population measure is also not without
limitations, the most noteworthy being omissions and biases related to OpenCellID users
themselves. Obviously, not everyone uses OpenCellID; the cell tower location data
reflect the movements of those who do. It would also not be a stretch to suggest that
OpenCellID users are probably younger than the average mobile phone user, given their
decision to use such an app. Still, the spatial patterns of cell density reflect local
knowledge about population centers and transportation corridors in Metro Vancouver.
The ambient population measure constructed from this data source likely provides a
better estimation of the population at risk than the residential population.
Two other limitations relate to the use of census data. First, it has been
suggested that census data does not directly measure social disorganization constructs
(Andresen, 2014). Rather, self-report data is necessary to adequately capture mediating
factors such as sparse local friendship networks (see Sampson & Groves, 1989;
Lowenkamp et al., 2003). A similar argument could be made for routine activity theory.
Variables such as median income only act as proxies for concepts like the number of
suitable targets in an area. A second limitation of census data concerns their link to
residents of spatial units like dissemination areas. Some of the results from the current
60
study were surprising, such as the negative relationship between the percentage of
males aged 15-24 and ambient population-based rates of theft of bicycle and other theft.
Past research has consistently linked young males with increased crime rates (Hirschi &
Gottfredson, 1983). This finding may be a question of where people live versus where
their routine activities take them. All census variables correspond to residential, not
ambient populations. This means that even when ambient crime rates are used in
regression models, the independent variables are still based on the residential
population. There is no practical solution to this problem, but it should be acknowledged.
This issue may explain the lower Pseudo R2 values for ambient population-based rate
models as a result of omitted variable bias. The ‘right’ independent variables that
correspond to sociodemographic and socioeconomic indicators for ambient populations
are unavailable.
6.4. Future Directions
Despite the limitations detailed above, the results from this study clearly
demonstrate the importance and value of considering the ambient population in crime
analysis. While the current research cannot say definitively whether this particular
measure of the ambient population provides a better estimation of the population at risk
than the residential population, it is clear that it impacts both spatial patterns and
regression results substantially. Future studies should make use of the ambient
population alongside the residential population, as the data are now easier than ever to
obtain (Andresen, 2006b). This and other ambient population measures should be
applied to different settings, at different spatial scales, and with disaggregated crime
data. Regarding the aggregation of crime data, many of the studies discussed in this
paper employed aggregate measures of crime (see Andresen 2006b, 2011; Kurland et
al., 2015; Mburu & Helbich, 2016). It is entirely possible that important trends are being
masked when various crime types are aggregated into a single measure. Overall, more
widespread use of the ambient population is recommended, particularly in a multivariate
context.
61
References
Andresen, M. A. (2006a). A spatial analysis of crime in Vancouver, British Columbia: A synthesis of social disorganization and routine activity theory. The Canadian Geographer, 50(4), 487-502.
Andresen, M. A. (2006b). Crime measures and the spatial analysis of criminal activity. The British Journal of Criminology, 46(2), 258-285.
Andresen, M. A. (2007). Location quotients, ambient populations, and the spatial analysis of crime in Vancouver, Canada. Environment and Planning A, 39(10), 2423-2444.
Andresen, M. A. (2011). The ambient population and crime analysis. The Professional Geographer, 63(2), 193-212.
Andresen, M. A. (2014). Environmental criminology. New York, NY: Routledge.
Andresen, M. A., & Brantingham, P. J. (2007). Hot spots of crime in Vancouver and their relationship with population characteristics. Ottawa, ON: Department of Justice Canada.
Andresen, M. A., & Jenion, G. W. (2008). Crime prevention and the science of where people are. Criminal Justice Policy Review, 19(2), 164-180.
Andresen, M. A., & Jenion, G. W. (2010). Ambient populations and the calculation of crime rates and risk. Security Journal, 23(2), 114-133.
Andresen, M. A., & Linning, S. J. (2012). The (in)appropriateness of aggregating across crime types. Applied Geography, 35(1-2), 275-282.
Andresen, M. A., & Malleson, N. (2013). Spatial heterogeneity in crime analysis. In M. Leitner (Ed.), Crime modeling and mapping using geospatial technologies (pp. 3-23). New York, NY: Springer.
Aultman-Hall, L., Lane, D., & Lambert, R. R. (2009). Assessing impact of weather and season on pedestrian traffic volumes. Transportation Research Journal: Journal of the Transportation Research Board, 2140, 35-43.
Blau, P. (1977). Inequality and heterogeneity. New York, NY: The Free Press.
Block, S., Clarke, R. V., Maxfield, M. G., & Petrossian, G. (2012). Estimating the number of U.S. vehicles stolen for export using crime location quotients. In M. A. Andresen & J. B. Kinney (Eds.), Patterns, prevention, and geometry of crime (pp. 54-68). New York, NY: Routledge.
62
Boggs, S. L. (1965). Urban crime patterns. American Sociological Review, 30(6), 899-908.
Bogomolov, A., Lepri, B., Staiano, J., Oliver, N., Pianesi, F., & Pentland, A. (2014). Once upon a crime: Towards crime prediction from demographics and mobile data. In ICMI ’14 Proceeding of the 16th International Conference on Multimodal Interaction (pp. 427-434).New York, NY, USA: ACM.
Boivin, R. (2013). On the use of crime rates. Canadian Journal of Criminology and Criminal Justice, 55(2), 263-277.
Boivin, R. (2018). Routine activity, population(s) and crime: Spatial heterogeneity and conflicting propositions about the neighborhood crime-population link. Applied Geography, 95, 79-87.
Boivin, R., & Felson, M. (2018). Crimes by visitors versus crimes by residents: The influence of visitor inflows. Journal of Quantitative Criminology, 34(2), 465-480.
Brantingham, P. J., & Brantingham, P. L. (1981). Introduction: The dimensions of crime. In P. J. Brantingham & P. L. Brantingham (Eds.), Environmental criminology (pp. 7-26). Prospect Heights, IL: Waveland Press.
Burgess, E. W. (1916). Juvenile delinquency in a small city. Journal of the American Institute of Criminal Law and Criminology, 6(5), 724-728. Retrieved from: http://www.jstor.org/stable/1133346
Burgess, E. W. (1925). The growth of the city: an introduction to a research project. In R. E. Park & E. W. Burgess (Eds.), The city: Suggestions for investigation of human behavior in the urban environment (pp. 47 – 62). Chicago, IL: University of Chicago Press.
Canadian Wireless Telecommunications Association. (n.d.). Facts & figures. Retrieved from https://www.cwta.ca/facts-figures/
Chainey, S., & Desyllas, J. (2008). Modelling pedestrian movement to measure on-street crime risk. In L. Liu & J. Eck (Eds.). Artificial crime analysis systems (pp. 71-91). Hershey, PA: IGI Global.
Chard, J., & Renaud, V. (1999). Visible minorities in Toronto, Vancouver and Montreal. Canadian Social Trends, 54, 20-25.
Clarke, R. V. (1984). Opportunity-based crime rates: The difficulties of further refinement. The British Journal of Criminology, 24(1), 74-83.
Cohen L. E., & Felson M. (1979). Social change and crime rate trends: A routine activity approach. American Sociological Review, 44, 588-608.
63
Cohen, L. E., Kaufman, R. L., & Gottfredson, M. R. (1985). Risk-based crime statistics: A forecasting comparison for burglary and auto theft. Journal of Criminal Justice, 13(5), 445-457.
Felson, M., & Boivin, R. (2015). Daily crime flows within a city. Crime Science, 4(1), 1-10.
Frank, J., Mannor, S., & Precup, D. (2013). Generating storylines from sensor data. Pervasive and Mobile Computing, 9(6), 838-847.
Gibbs, J. P., & Erickson, M. L. (1976). Crime rates of American cities in an ecological context. American Journal of Sociology, 82(3), 605-620.
Gerber, M. S. (2014). Predicting crime using Twitter and kernel density estimation. Decision Support Systems, 61, 115-125.
Glyde, J. (1856). Localities of crime in Suffolk. Journal of the Statistical Society of London, 19, 102-106. Retrieved from: http://www.jstor.org.proxy.lib.sfu.ca/stable/2338263
Guerry, A. M. (1833). Essai sur la statistique morale de la France. Paris: Crochard.
Hanaoka, K. (2018). New insights on relationships between street crimes and ambient population: Use of hourly population data estimated from mobile phone users’ locations. Environment and Planning B: Urban Analytics and City Science, 45(2), 295-311.
Harries, K. D. (1991). Alternative denominators in conventional crime rates. In P. J. Brantingham & P. L. Brantingham (Eds.). Environmental criminology (pp. 147-165). Prospect Heights, IL: Waveland Press, Inc.
Hewitt, A. N., Beauregard, E., Andresen, M. A., & Brantingham, P. L. (2018). Identifying the nature of risky places for sexual crime: The applicability of crime pattern and social disorganization theories in a Canadian context. Journal of Criminal Justice, 57, 35-46.
Hipp, J. R., Bates, C., Lichman, M., & Smyth, P. (2018). Using social media to measure temporal ambient population: Does it help explain local crime rates? Justice Quarterly. Advance online publication. doi: http://dx.doi.org/10.1080/07418825.2018.1445276
Hirschi, T., & Gottfredson, M. (1983). Age and the explanation of crime. American Journal of Sociology, 89(3), 552-584.
Hodler, R., & Raschky, P. A. (2017). Ethnic politics and the diffusion of mobile technology in Africa. Economics Letters, 159, 78-81.
64
Kadar, C., & Pletikosa, I. (2018). Mining large-scale human mobility data for long-term crime prediction. EPJ Data Science, 7(1), 1-27.
Kennedy, L. W., & Forde, D. R. (1990). Routine activities and crime: An analysis of victimization in Canada. Criminology, 28(1), 137-152.
Kounadi, O., Ristea, A., Leitner, M., & Langford, C. (2018). Population at risk: Using areal interpolation and Twitter messages to create population models for burglaries and robberies. Cartography and Geographic Information Science, 45(3), 205-220.
Kurland, J., Johnson, S. D., & Tilley, N. (2014). Offenses around stadiums: A natural experiment on crime attraction and generation. Journal of Research in Crime and Delinquency, 51(5), 5- 28.
Lee, C., Shih, C., & Chen, Y. (2013). Stochastic geometry based models for modelling cellular networks in urban areas. Wireless Networks, 19(6), 1063-1072.
Lee, M. (2017, November 27). Rising housing costs in Vancouver: New evidence from the census. Policynote. Retrieved from https://www.policynote.ca/rising-housing-costs-in-vancouver-new-evidence-from-the-census/
Ley, D., & Smith, H. (2000). Relations between deprivation and immigrant groups in large Canadian cities. Urban Studies, 37(1), 37-62.
Ley, D., & Murphy, P. (2001). Immigration in gateway cities: Sydney and Vancouver in comparative perspective. Progress in Planning, 55(3), 119-194.
Lottier, S. (1938). Distribution of criminal offenses in sectional regions. Journal of Criminal Law and Criminology, 29(3), 329-344.
Lowenkamp, C. T., Cullen, F. T., and Pratt, T. C. (2003). Replicating Sampson and Groves’s test of social disorganization theory: Revisiting a criminological classic. Journal of Research in Crime and Delinquency, 40 (4), 351-373.
MacDonald, Z. (2001). Revisiting the dark figure: A microeconometric analysis of under-reporting of property crime and its implications. The British Journal of Criminology, 41(1), 127-149.
Malleson, N., & Andresen, M. A. (2015a). Spatio-temporal crime hotspots and the ambient population. Crime Science, 4(1).
Malleson, N., & Andresen, M. A. (2015b). The impact of using social media data in crime rate calculations: Shifting hot spots and changing spatial patterns. Cartography and Geographic Information Science, 42(2).
Malleson, N., & Andresen, M. A. (2016). Exploring the impact of ambient population measures on London crime hotspots. Journal of Criminal Justice, 46, 52-63.
65
Mayhew, H. (1861). London labour and the London poor, volume IV: Those that will not work, comprising prostitutes, thieves, swindlers and beggars. London: Griffin-Bohn.
Mburu, L. W., & Helbich, M. (2016). Crime risk estimation with a commuter-harmonized ambient population. Annals of the American Association of Geographers, 106(4), 804-818.
Mletzko, D., Summers, L., & Arnio, A. N. (2018). Spatial patterns of urban sex trafficking. Journal of Criminal Justice, 58, 87-96.
More than half of Vancouver’s homeless population have been homeless for less than a year, count finds. (2018, May 1), CBC News. Retrieved from https://www.cbc.ca/news/canada/british-columbia/vancouver-homeless-count-2018-1.4644233
Murray, R. K., & Swatt, M. L. (2013). Disaggregating the relationship between schools and crime: A spatial analysis. Crime & Delinquency, 59(2), 163-190.
Nogueira de Melo, S., Pereira, D. V. S., Andresen, M. A., & Matias, L. F. (2017). Spatial/temporal variations of crime: A routine activity theory perspective. International Journal of Offender Therapy and Comparative Criminology, 62(7), 1967-1991.
OpenCellID. (2018). What is OpenCellID? Retrieved from http://wiki.opencellid.org/wiki/What_is_OpenCellID
Pereira, D. V. S., Mota, C. M. M., & Andresen, M. A. (2017). Social disorganization and homicide in Recife, Brazil. International Journal of Offender Therapy and Comparative Criminology, 61(14), 1570-1592.
Piza, E. L., & Gilchrist, A. M. (2018). Measuring the effect heterogeneity of police enforcement actions across spatial contexts. Journal of Criminal Justice, 54, 76-87.
Quetelet, L. A. J. (1842). A treatise on man and the development of his faculties. Edinburgh: W. and R. Chambers.
Rice, K. J., & Smith, W. R. (2002). Sociological models of automotive theft: Integrating routine activity and social disorganization approaches. Journal of Research in Crime and Delinquency, 39(3), 304-336.
Sampson, R. J., & Groves, W. B. (1989). Community structure and crime: Testing social-disorganization theory. American Journal of Sociology, 94(4), 774-802.
Shaw, C. R., & McKay, H. D. (1931). Social factors in juvenile delinquency. Washington, DC: U.S. Government Printing Office.
66
Shaw, C. R., & McKay, H. D. (1942). Juvenile delinquency and urban areas: A study of rates of delinquency in relation to differential characteristics of local communities in American cities. Chicago: University of Chicago Press.
Shaw, C. R., Zorbaugh, F., McKay, H. D., & Cottrell, L. S. (1929). Delinquency areas: A study of the geographic distribution of school truants, juvenile delinquents, and adult offenders in Chicago. Chicago, IL: University of Chicago Press.
Skogan, W. G. (1976). Victimization surveys and criminal justice planning. University of Cincinnati Law Review, 45(2), 167-206.
Smith, W. R., Frazee, S.G., & Davison, E. L. (2000). Furthering the integration of routine activity and social disorganization theories: Small units of analysis and the study of street robbery as a diffusion process. Criminology, 38(2), 489-523.
Sparks, R. F. (1980). Criminal opportunities and crime rates. In S. E. Fienberg & A. J. Reiss, Jr. (Eds.). Indicators of crime and criminal justice: Quantitative studies (pp. 18-28). U.S. Department of Justice, Bureau of Justice Statistics.
Statistics Canada. (2009). 2006 Census collection. Retrieved from http://www12.statcan.gc.ca/census-recensement/2006/ref/about-apropos/coll-eng.cfm
Statistics Canada. (2017). Police-reported crime statistics, 2016. Retrieved from https://www150.statcan.gc.ca/n1/daily-quotidien/170724/dq170724b-cansim-eng.htm
Stults, B. J., & Hasbrouck, M. (2015). The effect of commuting on city-level crime rates. Journal of Quantitative Criminology, 31(2), 331-350.
Traunmueller, M., Quattrone, G., & Capra, L. (2014). Mining mobile phone data to investigate urban crime theories at scale. In L. M. Aiello & D. McFarland (Eds.), Social Informatics: 6th International Conference, SocInfo 2014, Barcelona, Spain, November 11-13, 2014 Proceedings (pp. 396-411). New York, NY: Springer.
Vancouver. (n.d.) In Wikipedia. Retrieved August 27, 2018 from https://en.wikipedia.org/wiki/Vancouver
Willits, D., Broidy, L., & Denman, K. (2013). Schools, neighborhood risk factors, and crime. Crime & Delinquency, 59(2), 292-315.
Wooldredge, J. (2002). Examining the (ir)relevance of aggregation bias for multilevel studies of neighborhoods and crime with an example comparing census tracts to official neighborhoods in Cincinnati. Criminology, 40(3), 681-710.
Xie, L., Heegaard, P. E., & Jiang, Y. (2017). Survivability analysis of a two-tier infrastructure-based wireless network. Computer Networks, 128, 28-40.
